Deep Adversarial Sliced Wasserstein Domain Adaptation Neural Networks and Applications to Disaster Datasets

During a crisis, people resort to social media to share information that can be used by emergency responders and humanitarian organizations. Due to the amount of data that is gathered from social media, the task of retrieving relevant information becomes harder to do manually. Machine learning can speed up many human processes to attain rapid response for crisis management. One approach is to utilize a pre-trained model that recognizes damage-related images and filters out irrelevant ones without creating a new model each time a disaster occurs. However, a model trained on past disaster data do not always generalize well on data from a new disaster due to domain shift. Therefore, existing models need to be adapted to the emergent disaster quickly. A promising way to achieve this goals is through Unsupervised Domain Adaptation (UDA), which is one of the major fields of machine learning used when we need to train a model on unseen data from a different domain. To this end, many advancements have been made toward better transferability of the model, including the Domain Adversarial Neural Network (DANN) which utilized adversarial training to obtain a domain-invariant feature extractor. Adversarial Sliced-Wasserstein Domain Adaptation Network (AWDAN) further improved DANN by introducing a sliced-Wasserstein distance as a measure between the features extracted from the source and target domains. Inspired by these advancements, in this thesis, we explored the utility of UDA approaches in the disaster response domain. Specifically, we implemented and tested both DANN and AWDAN on a popular disaster damage assessment dataset, which comprises social media images from four natural disaster. Our experimental results and detailed analyses show that we can improve upon previous methods and accomplish state-of-the-art results with AWDAN for the disaster damage assessment task.